Trustworthy AI in Julia

The Alan Turing Institute

Patrick Altmeyer

Delft University of Technology

2024-05-08

Introduction

Economist by training, previously Bank of England, currently 3rd year PhD in Trustworthy AI @ TU Delft.

Motivation

Why Trustworthy AI and why in Julia?

  • Opaque AI technologies have entered the public domain with far-reaching stakes.
  • These technologies are here to stay, so at best, we can make them more trustworthy.
  • Julia has an edge:
    • Transparency: most packages are written in pure Julia.
    • Intuitiveness: great Lisp-like support for symbolic computing.
    • Community: welcoming, supportive and diverse (sort of!).
    • Autodiff: top-notch support, which helps with common XAI approaches.

Outline

  • Taija: A brief overview of the Taija ecosystem.
    • Overview, Projects, Research.
  • Deep Dive: A closer look at some of our core packages.
    • Counterfactual Explanations, Conformal Prediction, Laplace Redux, Joint Energy Models.
  • The Journey: Julia throught my PhD
    • From “I’ll try this out” to “I’ll never go back”.

Taija

Overview

Milestones

2021

  • First small-scale project in Julia on Bayesian regression.

Milestones

2022

  • Presented CounterfactualExplanations.jl and LaplaceRedux.jl at JuliaCon.

2021

  • First small-scale project in Julia on Bayesian regression.

Milestones

2023

  • Presented ConformalPrediction.jl at JuliaCon.
  • TU Delft students work on CounterfactualExplanations.jl and LaPlaceRedux.jl.
  • CounterfactualExplanations.jl published in JuliaCon proceedings.

2022

  • Presented CounterfactualExplanations.jl and LaplaceRedux.jl at JuliaCon.

2021

  • First small-scale project in Julia on Bayesian regression.

Milestones

2024

  • Multiple presentations at JuliaCon this summer.
  • GSoC/JSoC projects on Causal Counterfactuals and Conformal Bayes.
  • TU Delft students working on TaijaInteractive.jl.

2023

  • Presented ConformalPrediction.jl at JuliaCon.
  • TU Delft students work on CounterfactualExplanations.jl and LaPlaceRedux.jl.
  • CounterfactualExplanations.jl published in JuliaCon proceedings.

2022

  • Presented CounterfactualExplanations.jl and LaplaceRedux.jl at JuliaCon.

2021

  • First small-scale project in Julia on Bayesian regression.

Research

Taija has been used in the following publications:

  • Conformal Intent Classification and Clarification for Fast and Accurate Intent Recognition (Hengst et al. 2024) upcoming in ACL’s NAACL Findings 20241.
  • Faithful Model Explanations through Energy-Constrained Conformal Counterfactuals (Altmeyer et al. 2024) published in Proceedings of the AAAI Conference on Artificial Intelligence 2024.
  • Explaining Black-Box Models through Counterfactuals (Altmeyer, Deursen, et al. 2023) published in JuliaCon Proceedings.
  • Endogenous Macrodynamics in Algorithmic Recourse (Altmeyer et al. 2023) published in Proceedings of the 2023 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML).

Counterfactual Explanations

CounterfactualExplanations.jl: A package for Counterfactual Explanations and Algorithmic Recourse in Julia.

Conformal Prediction

ConformalPrediction.jl: Predictive Uncertainty Quantification through Conformal Prediction for Machine Learning models trained in MLJ.

Laplace Redux

LaplaceRedux.jl: Effortless Bayesian Deep Learning through Laplace Approximation for Flux.jl neural networks.

Joint Energy Models

JointEnergyModels.jl: A package for Joint Energy Models and Energy-Based Models in Julia.

The Journey

The Good

The Bad

Questions?

With thanks to my co-authors Andrew M. Demetriou, Antony Bartlett, and Cynthia C. S. Liem and to the audience for their attention.

References

Altmeyer, Patrick, Giovan Angela, Aleksander Buszydlik, Karol Dobiczek, Arie van Deursen, and Cynthia CS Liem. 2023. “Endogenous Macrodynamics in Algorithmic Recourse.” In 2023 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML), 418–31. IEEE.
Altmeyer, Patrick, Arie van Deursen, et al. 2023. “Explaining Black-Box Models Through Counterfactuals.” In Proceedings of the JuliaCon Conferences, 1:130. 1.
Altmeyer, Patrick, Mojtaba Farmanbar, Arie van Deursen, and Cynthia CS Liem. 2024. “Faithful Model Explanations Through Energy-Constrained Conformal Counterfactuals.” In Proceedings of the AAAI Conference on Artificial Intelligence, 38:10829–37. 10.
Hengst, Floris den, Ralf Wolter, Patrick Altmeyer, and Arda Kaygan. 2024. “Conformal Intent Classification and Clarification for Fast and Accurate Intent Recognition.” https://arxiv.org/abs/2403.18973.

Quote sources

  • “There! It’s sentient”—that engineer at Google (probably!)
  • “Algorithms don’t listen, nor do they bend”—Cathy O’Neil
  • “We’re fascinated with robots because they are reflections of ourselves.”—Ken Goldberg

Hiddens